How the translation telephone game works

Machine translation is optimised for accuracy in isolated, direct translation — treating a sentence in isolation and finding its closest equivalent in the target language. It does this job remarkably well. But language is not a simple look-up table. Words carry connotation, register, cultural reference, and metaphor that cannot always survive a border crossing.

When you pass text through a chain of translations, each step introduces what linguists call semantic drift — a small shift in meaning caused by imperfect equivalence. Compounded across six, eight, or ten languages, the drift accumulates until the final sentence bears only a family resemblance to the original — exactly like the whispered message in the schoolyard telephone game, just played with translation engines instead of classmates.

This is what makes Translation Mixer genuinely surprising even when you already know what to expect. The chain is non-deterministic in practice — the same sentence routed through a different set of languages will produce a completely different result. Every mix is unique.

Not all translation hops are equal, and the answer surprised us. We ran the same paragraph out and back through 20 languages, one at a time, and scored how much survived. The ranking below is our own measured data, not a guess:

For maximum chaos, put Japanese, Korean, or Chinese in your chain. For gentler, more poetic drift, stick to Romance or Germanic languages — Icelandic finished dead last for damage, and the word choices along the way are often beautifully unexpected. The full ranking, method, and damage scores are in We Round-Tripped One Paragraph Through 20 Languages.

Some text transforms far better than others. Here’s what tends to produce the most shareable, surprising, or hilarious results:

Translation Mixer is a custom-built web application. The frontend is built with React and Next.js, handling the language selection interface, result display, and user accounts. Under the hood, requests are proxied to a custom backend written in Go, which processes the chain of translation API calls sequentially, handles rate limiting, and stores saved translations for registered users.

Each translation in the chain is made through the Google Translate or Microsoft Azure Translator API. Translation Mixer orchestrates those API calls sequentially, feeding the output of each step as the input to the next, picking whichever provider supports that particular language pair, and collects the full path so you can see exactly which languages your text visited along the way.

We support over 200 languages between the two providers — including a few, like Upper and Lower Sorbian, Faroese and Klingon, that only one of them offers — so there are hundreds of thousands of possible chain combinations. No two mixes are ever quite the same.